fix(discovery): prevent premature node eviction from routing table

The replaceNode function removes nodes if either the replacement cache
has entries or the node's reliability is below the chosen threshold.

This means that nodes with perfect or good reliability (seen >= 0.5) are
evicted just because replacement candidates exist, completely ignoring
the reliability threshold.

Evidence from production logs show nodes with excellent reliability
being removed:

DBG - Node added to routing table           topics="discv5 routingtable" tid=1 n=1ff*7a561e:10.244.0.208:6890
DBG - bucket                                topics="discv5" tid=1 depth=0 len=2 standby=0
DBG - node                                  topics="discv5" tid=1 n=130*db8a1b:10.244.2.207:6890 rttMin=1 rttAvg=2 reliability=1.0
DBG - node                                  topics="discv5" tid=1 n=1ff*7a561e:10.244.0.208:6890 rttMin=1 rttAvg=14 reliability=1.0
DBG - Node removed from routing table       topics="discv5 routingtable" tid=1 n=1ff*7a561e:10.244.0.208:6890
DBG - Total nodes in discv5 routing table   topics="discv5" tid=1 total=1
DBG - bucket                                topics="discv5" tid=1 depth=0 len=1 standby=0
DBG - node                                  topics="discv5" tid=1 n=130*db8a1b:10.244.2.207:6890 rttMin=1 rttAvg=165 reliability=0.957
DBG - Node removed from routing table       topics="discv5 routingtable" tid=1 n=130*db8a1b:10.244.2.207:6890
DBG - Total nodes in discv5 routing table   topics="discv5" tid=1 total=0

First entry shows a node with perfect reliability (1.0) and 14ms RTT
being removed. Second one shows a node with 95.7% reliability and
minimal RTT also being evicted. Both far exceed 0.5 threshold set by
NoreplyRemoveThreshold.

Nodes are now only removed if their reliability actually falls below the
specified threshold, regardless of replacement cache status. This allows
the reliability tracking to properly tolerate transient failures.

The replacement cache logic inside the conditional remains unchanged,
if a replacement exists, it's used; otherwise the node is just removed.

Signed-off-by: Chrysostomos Nanakos <[email protected]>

Author: cnanakos

codexdht/private/eth/p2p/discoveryv5/routing_table.nim

@@ -445,7 +445,7 @@ proc replaceNode*(r: var RoutingTable, n: Node, forceRemoveBelow = 1.0) =
   ## also avoids being too agressive because UDP losses or temporary network
   ## failures.
   let b = r.bucketForNode(n.id)
-  if (b.replacementCache.len > 0 or n.seen <= forceRemoveBelow):
+  if n.seen <= forceRemoveBelow:
     if b.remove(n):
       debug "Node removed from routing table", n
       ipLimitDec(r, b, n)